A startup whose product competes with GitHub Copilot and different AI-powered coding assistants has achieved unicorn standing.
On Thursday, Codeium mentioned it closed a $150 million Collection C spherical led by Basic Catalyst that values the corporate at $1.25 billion post-money. The spherical, which additionally noticed participation from present buyers Kleiner Perkins and Greenoaks Capital, brings the corporate’s complete funding raised to just about a quarter-billion {dollars} ($243 million) a mere three years since its launch.
Codeium’s co-founder and CEO, Varun Mohan, advised TechCrunch that Codeium hasn’t even touched the $65 million Collection B tranche it raised in January but. Again then, simply eight months in the past, Codeium was valued at half-a-billion {dollars}.
“Regardless that we’ve barely made a dent in our present funding, we consider that this injection of capital will enable us to considerably ramp up R&D and development whereas making even bigger strategic bets,” he mentioned.
Codeium was based in 2021 by Mohan and his childhood good friend and fellow MIT grad, Douglas Chen. Previous to Codeium, Chen was at Meta, the place he helped to construct software program instruments for VR headsets just like the Oculus Quest. Mohan was a tech lead at Nuro, the autonomous supply startup, liable for managing the autonomy infrastructure staff.
The startup started as a radically totally different firm known as Exafunction, centered on GPU optimization and virtualization for AI workloads. However in 2022, Mohan and Chen sensed a much bigger alternative in generative coding, and determined to rebrand — and pivot.
“Regardless of the inflow of generative AI instruments, builders are nonetheless battling time-consuming coding duties,” Mohan mentioned. “Lots of the AI-driven options present generic code snippets that require important handbook work to combine and safe inside present codebases. That’s the place our AI coding help is available in.“
Codeium’s platform, powered by generative AI fashions skilled on public code, serves up strategies within the context of an app’s total codebase. It helps round 70 programming languages and integrates with a lot of in style improvement environments, together with Microsoft Visible Studio and JetBrains.
To draw devs away from Copilot and different rivals, Codeium has launched a beneficiant free tier to begin. The technique appears to have labored: Immediately, the startup has greater than 700,000 customers and over 1,000 enterprise clients, together with Anduril, Zillow, Dell and AthenaHealth.
Quentin Clark, managing director at Basic Catalyst, implied that Codeium received a few of its bigger contracts by embracing a steadfastly client-centric method to product analysis.
“The staff’s method has at all times been to comply with its clients, main the corporate to construct options on their phrases — deployable in any atmosphere and supporting extra languages than anybody else,” Clark mentioned in an announcement. “What Codeium has created isn’t only a demo, an announcement, or an thought — this can be a fully-scaling enterprise, with giant enterprises adopting the product throughout their total organizations.”
Companies are sometimes cautious of exposing proprietary code to a 3rd occasion — as an illustration, Apple reportedly banned workers from utilizing Copilot final yr, citing issues about confidential information leakage. To try to allay such fears, Codeium started providing a self-hosted set up choice alongside its normal software-as-a-service plan.
Firms can now deploy Codeium’s service on their very own {hardware} if they need. Or they’ll undertake a hybrid setup, maintaining their information on their very own units whereas utilizing Codeium’s servers for computing wants.
There’s at all times some danger concerned in information transfers to the cloud, however Mohan claimed that Codeium leverages sturdy encryption. “We by no means practice our proprietary generative autocomplete mannequin on consumer information, by no means promote information and guarantee all information transmission is encrypted,” he added.
Codeium has additionally taken steps to take away “non-permissively” licensed code (e.g., code below copyright) from the info units it used to coach its AI fashions. Some code-generating instruments skilled utilizing restrictively licensed or copyrighted code have been proven to regurgitate that code when prompted in a sure method, posing a legal responsibility danger (builders that incorporate the code could possibly be sued). Mohan mentioned that’s not the case with Codeium, because of its coaching information prep and filtering method.
“We additionally take away any remaining information that appears much like code that’s explicitly non-permissively licensed simply in case different folks copied code with out offering the right attribution and licensing,” he added. “On high of this, we now have state-of-the-art, post-generation attribution filtering and logging within the case that these giant probabilistic fashions produce code that’s much like public code, whether or not permissively or non-permissively licensed.”
However what about hallucinations? Most AI coding instruments are infamous for making stuff up, which may be fairly damaging in an enterprise atmosphere.
An evaluation by developer tooling startup GitClear discovered that generative AI instruments have resulted in extra mistaken code being pushed to codebases over the previous few years. And a Purdue research discovered that over half the solutions that OpenAI’s ChatGPT provides to programming questions are incorrect. Safety researchers have warned of the potential for such instruments to amplify present bugs in software program.
A current survey from cybersecurity agency Synk discovered that 9 in ten builders fear concerning the broader safety implications of utilizing AI coding platforms. However Mohan claimed that Codeium’s supposedly superior, deep context-rich tech yields extra reliable outcomes than most.
“Our context consciousness engine is ready to floor ends in what’s already present in a consumer’s codebase, resulting in strategies with fewer hallucinations and extra adherence to present syntax, semantics and requirements,” he mentioned.
Whether or not benchmarks again that up or not, Codeium’s gross sales pitch appears to be resonating with the precise execs: Income hit eight figures this yr. Mohan mentioned the 80-person, San Jose-based startup plans to increase headcount to 120 by 2025 because it goals to make a much bigger dent in a market with formidable rivals like Tabnine, Anysphere and Poolside.
Catching as much as Copilot, which had over 1.3 million paying customers as of April, most likely isn’t within the playing cards for Codeium — no less than not imminently. It doesn’t should be. As Mohan rightly famous, given the widespread adoption of AI coding instruments amongst builders (regardless of their reservations), even a small slice of the nascent section is sure to be profitable.
Polaris Analysis initiatives that the AI code instruments market shall be price $27.17 billion by 2032.
“An overabundance of hype is a problem the trade faces,” Mohan mentioned. “It will make it tougher for each firm to really persuade finish customers that they’re on the forefront of chance. However we consider that truth-seeking and practical AI firms like Codeium will ultimately lower by means of this noise.”
